Fetching from the wire…
01
02
03
04
05
06
07
08
Models2026-08-18 · source-backed
An Aug 17 writeup benchmarks it on an RTX PRO 4000 Blackwell SFF using a custom 5.01 BPW iMatrix/NVFP4 hybrid quant (bulk matrices NVFP4, sensitive tensors Q5_K/Q6_K) and a patched llama.cpp with Gated DeltaNet support and chained MTP speculative decoding. Production mean 50.44 tok/s versus 21.19 target-only greedy, a 2.81x speculative speedup, validated at 261,500 input tokens, occupying 23,952 of 24,467 MiB. The honest caveats: throughput collapses to 12.61 tok/s at full cache, the max-TPS profile isn't bitwise distribution-preserving, and a higher-precision draft model actually cut throughput 26.6%.
Each link below shares sources, entities, or timing with this story.
NVIDIA released Blackwell / Shared entities / Shared topic / Earlier coverage / Downstream implication
Linked by a graph relationship (NVIDIA released Blackwell); both cover Blackwell, Qwen3, RTX PRO; overlapping topics (card, context).
NVIDIA released Blackwell / Shared entity: Blackwell / Earlier coverage / Tension
Linked by a graph relationship (NVIDIA released Blackwell); both cover Blackwell; earlier Blackwell coverage from 2026-06-02.
NVIDIA released Blackwell / Shared entities / Shared topic / Earlier coverage
Linked by a graph relationship (NVIDIA released Blackwell); both cover NVFP4, Qwen3; overlapping topics (context, full).
Linked by a graph relationship (NVIDIA released Blackwell); both cover Blackwell, Qwen3; overlapping topics (benchmark, blackwell).
NVIDIA released Blackwell / Shared entities / Earlier coverage / Tension
Linked by a graph relationship (NVIDIA released Blackwell); both cover Blackwell, RTX PRO; earlier Blackwell coverage from 2026-06-25.
Shared entities / Shared topic / Earlier coverage
Both cover Blackwell, Qwen3, TPS; overlapping topics (cache, throughput); earlier Blackwell coverage from 2026-08-09.
Both cover Blackwell, Qwen3, RTX PRO; overlapping topics (blackwell, card); earlier Blackwell coverage from 2026-07-24.
Both cover Blackwell, Qwen3, RTX PRO; overlapping topics (blackwell, custom); earlier Blackwell coverage from 2026-03-14.